Research On Handwriting Recognition And Its Related Problems

As important tools of information representation and communication,characters are widely used in people's daily life and work.Moreover,handwriting also is an essen-tial skill for everyone to survive in our society.With the development and popularization of computers,handwriting recognition as an advanced human-computer interaction way has attracted extensive attention in recent decades.Although researchers have made many great achievements in handwriting recognition,there still leaves a big room to improve deep learning based algorithms for handwriting recognition.Therefore,our primary objective is to propose more accurate and efficient handwriting recognition al-gorithms.Moreover,with the advances in sensing technology,a novel human-computer interaction way called in-air handwriting has been proposed,which allows users to write freely in midair with their own hands or fingers.However,the research on in-air handwriting recognition is still in its infancy,especially for in-air handwritten Chinese text recognition.Therefore,in-air handwriting recognition is our another research topic.Our main contributions are summarized as follows:1.We proposed a novel 1-dimensional convolutional neural network(1D-CNN)based algorithm for online handwritten Chinese character recognition(OLHCCR)from a new perspective,which utilizes the 1D-CNN to directly deal with the sequential structures of handwriting trajectories.The proposed method is completely different from existing methods for OLHCCR including the 2D-CNN or RNN(Recurrent Neu-ral Network)based methods,and it also enjoys several obvious advantages including:(1)Compared with the 2D-CNN,the 1D-CNN not only avoids well-studied domain-specific knowledge for extracting complicated directional feature maps,but it also achieves higher recognition accuracy with a more compact model;(2)Compared with the RNN,the 1D-CNN has faster computational speed when processing long handwrit-ing trajectories since its computation over the time dimension can be fully parallelized.2.We proposed an attention-based algorithm for handwritten English word recog-nition,which follows an encoder-decoder architecture to literally output target characters by integrating the attention mechanism.Moreover,we further improve this architec-ture with the following strategies including:(1)We adopted the 1D-CNN instead of the RNN for encoding handwriting trajectories,which largely accelerates the encoding speed;(2)We adopted the lexicon-based decoding,which significantly improves the recognition accuracy with integrating the linguistic contexts.Furthermore,we con-structed the first large-scale in-air handwritten English word(IAHEW)dataset called IAHEW-UCAS2016 to promote future research on IAHEW recognition.Finally,we have built a practical large-vocabulary IAHEW recognition system.3.We proposed a novel temporal convolutional recurrent network(TCRN)for online handwritten Chinese text recognition(OLHCTR).The proposed method enjoys several advantages over existing methods including:(1)Compared with the prevalent RNN,the TCRN can achieve higher recognition accuracy with faster prediction;(2)Compared with the state-of-art method(i.e.,the fully convolutional recurrent network),the TCRN not only avoids the well-studied domain-specific knowledge for complicated feature map extraction,but it also attains higher training efficiency(including lower computation complexity,less memory consumption,faster convergence)with a more compact model.Furthermore,we constructed the first 3D in-air handwritten Chinese text(IAHCT)recognition system in the world.As far as we know,there is no research on IAHCT recognition before.Therefore,we constructed the first large-scale IAHCT dataset called IAHCT-UCAS2018 to promote future research on IAHCT recognition.4.We proposed a novel handwriting imitation generative adversarial network(Hi-GAN)for handwritten text synthesis.We find that existing handwriting generative models have not entirely solved the handwritten text generation,since they are limited to synthesizing either short handwritten words or long handwritten texts with random styles.Instead,the proposed HiGAN can generate variable-length handwritten texts conditioned on arbitrary textual contents,which are unconstrained to any pre-defined corpus or out-of-vocabulary words.Moreover,HiGAN can exactly disentangle the cal-ligraphic styles of reference samples and further imitate generating handwriting images of similar styles.Our experimental analysis demonstrates that HiGAN shows satisfied generalization for handwriting synthesis rather than memorizing the training data.5.For handwritten Chinese character recognition(HCCR)models based on CNNs,we proposed a novel unified network compression algorithm.The proposed compres-sion algorithm integrates the advantages of channel pruning and network quantization,which can fully compress the model with no significant accuracy loss.Specifically,for channel pruning,we adopted a coarsen-to-fine strategy to gradually prune the redundant channels of the CNN;for network quantization,we introduced extra drop-weights and incremental quantization strategies to convert original full-precision weights into low-bit representation as far as possible.Experiments on public handwriting datasets(includ-ing ICDAR-2013,IAHCC-UCAS2016&MNIST)demonstrate that our algorithm can significantly compress CNN-based models for HCCR.